[TASK] Refactor background-job liveness monitoring onto the Supervisor/BackgroundJob contract

[edenreich]

Summary

Reframed from the original bug report (kept below): the broken A2A_QueryTask guard turned out to be one symptom of fragmented background-job liveness monitoring. Since #693 all three background job kinds - A2A tasks, background shells, subagents - implement the uniform BackgroundJob contract and run under the job supervisor, but the readers/controllers still consulted different sources:

Consumer	A2A	Shell	Subagent
Status bar / /tasks list	supervisor	supervisor	supervisor
HasPending()	supervisor	ShellTracker	SubagentTracker
A2A_QueryTask polling guard	tracker, wrong key (silent no-op)	none	none
/tasks Cancel	works	A2A-only path, silently fails	silently fails

Scope

1. Add a uniform per-id liveness query to the shared contract: Supervisor.IsRunning(id), exposed as BackgroundTaskRegistry.IsJobRunning(id) and consumed by tools via a narrow JobLivenessReporter projection (mirroring JobSubmitter/JobStopper).
2. Fix the A2A_QueryTask guard as its first consumer - keyed by task id, sourced from the supervisor (the original bug).
3. Add JobMeta.HoldsSession so each job declares whether it keeps a headless session alive; Supervisor.HasPending() reads it and the registry delegates - shells and subagents stop consulting their trackers for liveness, and the interactive-pane carve-out is declared by the job itself instead of a tracker-side mode filter.
4. Fix /tasks Cancel for shells and subagents: the confirm gate arms for running non-A2A rows and cancellation dispatches by kind - A2A keeps CancelBackgroundTask (remote cancel + tracker cleanup; a2aJob.Wind is deliberately a local no-op), shells/subagents wind their supervised job with WindJob(id, WindStop).

Follow-up from #693.

Original bug report

A2A_QueryTask passes an agent URL where a task ID is expected. In internal/agent/tools/a2a_query_task.go the "is this task still being polled" guard calls IsPolling(agentURL) / GetPollingState(agentURL), but A2ATaskTracker.IsPolling/GetPollingState are keyed by task ID (taskIndex), so the lookup never matches and the guard is a silent no-op.

Steps to Reproduce

1. Submit a long-running A2A task via A2A_SubmitTask.
2. Call A2A_QueryTask for that task while it is polling.
3. Inspect the "polling active" field of the result.

Expected Behavior

A2A_QueryTask reports the task as actively polling while it is in flight.

Actual Behavior

The polling-active guard never matches (agent URL is passed where a task ID is expected), so it always reports not-polling regardless of the task's real state.